Академический Документы
Профессиональный Документы
Культура Документы
Interfaces
Fabien Lotte, Laurent Bougrain, Maureen Clerc
Electroencephalography (EEG)-based
Brain-Computer Interfaces
Fabien LOTTE1 , Laurent BOUGRAIN2 , Maureen CLERC3
1
Inria Bordeaux Sud-Ouest, FRANCE
2
Lorraine University/Inria Nancy Grand-Est, FRANCE
3
Inria Sophia Antipolis Mediterranee, FRANCE
June 1, 2015
Abstract
Brain-Computer Interfaces (BCI) are systems that can translate the
brain activity patterns of a user into messages or commands for an interactive application. The brain activity which is processed by the BCI systems
is usually measured using Electroencephalography (EEG). In this article,
we aim at providing an accessible and up-to-date overview of EEG-based
BCI, with a main focus on its engineering aspects. We notably introduce some basic neuroscience background, and explain how to design an
EEG-based BCI, in particular reviewing which signal processing, machine
learning, software and hardware tools to use. We present Brain Computer
Interface applications, highlight some limitations of current systems and
suggest some perspectives for the field.
Introduction
technology [11, 6]. BCI have indeed evolved to become more robust, more accessible and promising for a wide range of applications including communication
and control for motor impaired users [9, 12], gaming targeted toward the general
public [13, 14], real-time mental state monitoring [15] or stroke rehabilitation
[16, 17].
Designing a BCI is a complex and difficult task which requires multidisciplinary knowledge in computer science, engineering, signal processing, neuroscience and psychology. In order to use a BCI, two phases are generally required:
1) an offline training phase during which the system is calibrated and 2) the
operational online phase in which the system recognises brain activity patterns
and translates them into commands for a computer. An online BCI system is a
closed-loop, generally composed of six main steps: brain activity measurement,
preprocessing, feature extraction, classification, translation into a command and
feedback:
1. Brain activity measurement allows to acquire the raw signals reflecting
the users brain activity [18]. Various types of sensors can be employed,
but in this article we focus on EEG as the measurement technique.
2. Preprocessing consists in cleaning and denoising input data in order to
enhance the relevant information contained in the raw signals [19].
3. Feature extraction aims to describe the signals by a few relevant values
called features [19]. These features can be, for instance, the power of
the EEG over selected channels, and in specific frequency bands.
4. Classification assigns a class to a set of features extracted from the
signals within a certain time range [20]. This class corresponds to the kind
of brain activity pattern identified (e.g., imagined left hand movement or
imagined right hand movement). Classification algorithms are known as
classifiers.
5. Translation into a command associates a command to the brain activity pattern identified in the users brain signals, e.g., a recognised left
hand movement could be translated into the command move the cursor
left. This command is then used to control a given application such as a
speller (text editor, see part 4.1) or a robot [21].
6. Feedback is provided to the user to inform him/her about the recognised
brain activity pattern. This aims to help the user modulate his/her brain
activity and as such improve his/her control over the BCI [5]. Indeed,
BCI is a skill that needs to be learned and refined [22].
This whole architecture is summarised in Figure 1. These steps define an
online BCI. The BCI research community is in quest of solutions to avoid a
time-consuming calibration stage before operating a BCI (see, e.g., [23, 24]).
Currently, calibration is generally necessary in order to obtain a reliable BCI
operation, and is generally done offline. In this stage, the classification algorithm
2
is calibrated and the optimal features, and relevant sensors are selected. For this
calibration, a training data set needs to be prerecorded from the user. Indeed,
EEG signals are highly user-specific, and as such, most current BCI systems are
calibrated specifically for each user. This training data set should contain EEG
signals recorded while the user performed each mental task of interest several
times, according to given instructions. The recorded EEG signals will be used
as mental state examples in order to find the best calibration parameters for
this user.
EEG-based BCI.
In this article, we aim at providing an accessible and up-to-date overview of
EEG-based BCI (EEG being by far the most used modality for BCI design), with
a main focus on its engineering aspects. After a brief presentation of some basic
BCI-related neuroscience background (Section 2), Section 3 describes how to
design an EEG-based BCI, and notably how to process and classify EEG signals.
Then, Section 4 presents the main current applications of BCI technologies,
while Section 5 details the main hardware and software components available
to design these BCI. Finally, Section 6 concludes the article with a discussion
on the future of BCI research and technology.
At the centre of the BCI loop is the user and his/her brain, whose activity is
being measured and processed by the BCI. In this section, we provide some brief
notion of brain anatomy relevant for BCI and briefly describe how to measure
brain activity and what kind of brain activity patterns can be detected and used
by the BCI.
2.1
The brain includes three main parts: the cerebellum, the brainstem and the
cerebrum (which includes the cortex). The cortex is composed of the right
and left cerebral hemispheres, each of which being subdivided into four lobes:
frontal, parietal, temporal and occipital. The central fissure (or central sulcus)
separates the parietal lobe from the frontal lobe, and the primary motor cortex
from the primary somatosensory cortex. The frontal lobe is associated with
motor learning, problem solving, language, reasoning, planning, memory, and
is the core for personality and emotions. The primary motor cortex, located in
this lobe, is associated with motor activity, for planning and executing movements. Broadly speaking, the muscles to the left of the body are controlled by
the right hemisphere, and the left hemisphere controls the right side. The primary motor cortex sends information needed to perform voluntary movements
and receive sensory information from other brain areas. The received sensory
information is used as feedback to plan and execute movements. Furthermore,
the primary motor cortex contains a broad representation of the anatomy, see
Figure 2, where different body parts are partially controlled by overlapping regions of the motor cortex.
The parietal lobe plays an important role in the perception of tactile stimuli: it responds to sensory inputs, such as temperature, pressure and pain. It is
also associated with movement coordination and language. The occipital lobe is
responsible for the early stages of processing of visual stimuli and spatial recognition. The temporal lobe plays an important role in the perception of auditory
and olfactory stimuli. It is also associated with coordination and balance, emo-
Figure 2: Primary motor cortex of the right hemisphere: point-by-point correspondence of a body area with a specific area of the primary motor cortex.
The central rectangle of Figure 3 indicates the corresponding locations on EEG
electrodes.
There are several systems for recording brain activity. They can be classified i) by the location at which the signal is measured (invasive vs non-invasive)
and ii) by the type of sensors used to record the brain signals. These sensors can measure electric, magnetic or chemical signals. BCI based on invasive
recordings such as electrocorticography and intracortical electrodes are mainly
performed with animals and human patients implanted for clinical diagnosis.
Non-invasive technologies, which do not penetrate the skin, do not represent
any risk for the user, such as, haemorrhages, infections or bioincompatibility.
Some of the available non-invasive techniques include functional magnetic resonance imaging (fMRI) and functional near infrared spectroscopy (fNIRS) - sensitive to blood oxygenation - positron emission tomography (PET) and single
photon emission computed tomography (SPECT) - based on nuclear medicine
- , transcranial Doppler ultrasound (TcD) - based on ultrasound - and finally
magnetoencephalography (MEG) and electroencephalography (EEG) - directly
measuring electromagnetic fields generated by neuronal electrical activity.
EEG is the most common method for recording brain signals in humans
because it is safe, affordable, easy to use, and because it has a high temporal
resolution (of the order of milliseconds) [34]. EEG electrodes, placed on the
scalp, can be either passive or active. Passive electrodes, which are metallic, are connected to an amplifier by a cable. Active electrodes have an inbuilt
preamplifier to make them less sensitive to environmental noise and cable movements. Both types of electrodes need gel or saline liquid to operate, in order
5
to reduce the skin-electrode contact impedance. This prevents their use for
long periods, such as a whole day, because the gel dries with time, causing the
impedances to rise and the signals to degrade. For this reason, dry electrodes
are currently being developed and tested for BCI (see Section 5). Despite a low
spatial resolution (at best the inter-distance between electrodes is 1-2 cm) and a
low signal-to-noise ratio, this solution presents many advantages for widespread
use.
In early recordings of brain activity using EEG, there was no rule concerning
the layout of the electrodes on the head. Then common protocols were defined to
standardise the layout, the first one being the 10-20 system [35], still used today.
The name of the system comes from the inter-electrode distances, which are
approximately 10% or 20% of the front-to-back or right-to-left head perimeters
(see Figure 3).
An amplifier is necessary for acquiring measurable signals. Modern amplifiers produce a set of digital signals, one per channel. Each channel measures
the electric potential difference between the electrode to which it is connected
and a reference electrode. This reference electrode can be placed anywhere on
the head, but should not be at a location where the activity is very strong,
because it could prevent weaker activity of other regions from being correctly
measured. Many positions can be selected as reference, for instance an earlobe,
or the mastoid region behind the ear. Some amplifiers apply a common average reference, by subtracting from each channel the average potential over all
channels, a good strategy when recording from a large number of electrodes (so
called high-density EEG). A ground electrode is also necessary to prevent power
line interference with the low-amplitude signals of interest. This ground can be
placed anywhere on the body, and it is sometimes also placed on the scalp.
The next sections describe the three main brain activity patterns that are used
to design EEG-based BCI: Event Related Desynchronization/Synchronization,
Event Related Potentials and Steady State Evoked Potentials. Note that these
are not the only existing brain activity patterns used in EEG-based BCI, but
only the most popular and so far the most efficient for communication and
control applications.
2.2
Figure 3: Extended 10-20 EEG system. Electrodes are identified with labels:
the letters correspond to cortical locations: frontal (F), temporal (T), parietal
(P), occipital (O), central (C) and combinations of these to indicate intermediate
locations (Fp means frontal polar and corresponds to the anterior pre-frontal
lobe). The electrode numbers indicate laterality: odd numbers to the left, even
numbers to the right, and the letter z (meaning zero) on the midline. The
electrode layout displayed considers 10% intermediate distances. The central
rectangle indicates the location of the primary motor cortex (adapted with the
permission of gtec medical engineering GmbH).
and ERS have been suggested to reflect cortical activation and deactivation,
respectively [36].
ERD/ERS can be observed during a number of different mental tasks, and
can thus be used as a feature to drive a BCI. In particular, ERD/ERS can be
observed during Motor Imagery (MI), as well as executed movement [37] (see,
e.g., Figure 7). For instance, an execution or an imagination of a left-hand
movement leads to an ERD in the (8-12Hz) and (16-24Hz) bands (the socalled sensorimotor rhythms) during movement, and to an ERS in the band
after movement. Both the ERD and the ERS occur in the controlateral sensorimotor cortex (in the hemisphere opposite to the hand used). ERD/ERS can
also be observed during diverse mental imagery tasks, such as mental rotation
of geometric figures or mental subtractions, among others [38, 39].
ERD/ERS-based BCI have three main advantages: 1) they can be rather
natural and intuitive to use, since their users just have to imagine a specific
task to send control commands, 2) they do not require any stimulus (contrary
to ERP-based BCI, see Section 2.3), which enables the users to devote all their
sensory attention to the feedback of the BCI and/or to their external environment 3) they can be used in a self-paced way, that is, the user can initiate the
task (and hence the command) at will. Their main drawback is the fact that
such BCI require substantial training on the user side, to find the correct imagination strategy that leads to the clearest and most distinct ERD/ERS. Also,
this kind of BCI can currently only be used with a maximum of 3 or 4 different
mental imagery tasks to ensure maximal performances: with more tasks, their
recognition accuracy becomes considerably degraded, and therefore the overall
BCI efficiency decreased [40].
2.3
Event Related Potentials (ERPs) are stereotyped deflections of the EEG, related
to a sensory, cognitive or motor event. The Event of an ERP must come at
a precise measurable time, called the time trigger of the event. Typically, if a
person is receiving a sensory stimulus, ths time trigger is the instant at which
the stimulus is presented, or, in case of a button press, it is the instant at which
the muscular activity exceeds a certain threshold. The deflections measured by
the ERPs are represented as waveforms, and sometimes called waves. But in
spite of this terminology, ERPs do not display sustained oscillations, contrarily
to the ERD/ERS of part 2.2.
ERP nomenclatures are very varied: they can indicate their context of occurrence (Somatosensory Evoked potential (SEP) - in response to a tactile stimulation; Error Potential (ErrP) - related to an error), or the polarity of the
deflection: negativity indicating a decrease in amplitude, and positivity an increase in amplitude. For instance, the Mismatch Negativity (MMN) - is a negative deflection of the EEG in response to a deviant stimulus within a sequence
of repeating stimuli, either auditory or visual [41]. The timing of the deflection
is indicative of the level of processing involved: very early components (20 to
100 ms after the event trigger) are automatic, reflex responses whereas later
deflections occurring after 200 ms, reflect a higher-level, cognitive, processing.
This timing is also often part of ERP nomenclatures: an auditory N100 or
visual N1 are early negative deflections that may be modulated by attention;
the P300 wave - positive deflection, occurring approximately 300 ms after the
presentation of a deviant stimulus to which the person is paying attention.
few tens of V), and are dominated by background activity, which makes them
difficult to detect; 2) There are large inter-individual differences in ERP latencies and waveforms, requiring the system to be trained to recognize the ERP
of a given individual. 3) ERPs are locked to an event occurring at a specific
time, so ideally the BCI protocol should provide this type of stimulation, while
controlling its precise timing.
2.4
Steady State Evoked Potentials (SSEP) are measured when the subject perceives a periodic stimulus such as a flickering picture or an amplitude-modulated
sound. SSEP are defined by an increase of the EEG power at the frequency of
the stimulus (e.g., the flickering frequency of the picture or the modulation
frequency of the sound) or at its harmonics or sub-harmonics [46, 47, 27]. Interestingly, SSEP are modulated by attention: the more attention the user pays
to a stimulus the higher the SSEP response. As such, if the user is provided
with multiple periodic stimulus, each stimulus with a different stimulation frequency, it is possible to detect in the EEG signals which of these stimuli the
user is attending to, by finding the largest SSEP response. Each stimulus can
be associated to a command, leading to SSEP-based BCI with multiple commands. Various kinds of SSEP are used for BCI, such as Steady State Visual
Evoked Potentials (SSVEP) [46, 47], somatosensory SSEP [48] or auditory SSEP
[49]. These SSEP can be observed over the brain areas corresponding to the
sensory channel being stimulated, such as the visual areas (occipital cortex)
when SSVEP are used. SSVEP are by far the most used SSEP for BCI design,
and were successfully used in multiple applications such as SSVEP-based BCI
wheelchair control [50], spellers [51] or gaming applications [52], among many
others.
SSVEP are increasingly popular as they can be used quickly without user
training, and can provide a large number of commands. Their main limitations
are the need for a stimulus, and the reliance on the users visual attention.
2.5
Hybrid BCI
Interestingly, all the different types of EEG patterns that we have just presented,
and that can be used to operate a BCI, do not require to be used in isolation.
Multiple such signals can be combined together, e.g., to control a wheelchair
based on both P300 and SSVEP [50], or combined with other control signals not
originating from the brain, such as eye trackers or joysticks [53, 54]. Such BCI
combining multiple types of control signals are known as Hybrid BCI [55]. Such
hybrid BCI approaches are a promising approach to make BCI more robust, by
combining multiple inputs and control signals, as well as to extend the target
user population of BCI. For instance, by combining EEG signals with signals
from muscular activity (electromyogram - EMG), such hybrid BCI could prove
useful not only for completely or severely paralysed users (historically, the initial
target population of BCI) but also for motor-impaired users with residual muscle
10
control, e.g. stroke patients. The interested reader can refer to the survey in
[56] for more information on the use of Hybrid BCI for clinical applications.
Once the EEG signals acquired, they must be analysed and processed in order
to identify the brain activity patterns they contain. As mentioned above, this
is achieved by first preprocessing the signals, then extracting features which are
then classified. These three steps are detailed below, and followed by a section
on sensory feedback, closing the BCI system loop.
3.1
Preprocessing
The electrical activity of the brain causes variations of electric potential on the
scalp of only a tens of few microVolts. Care must be taken when acquiring the
EEG, and when processing the signals, in order for the relevant information to
be preserved and revealed.
Part of the EEG signals comes from electrical activity unrelated to the brain:
muscular activity of the head and neck, eye movements which are equivalent to
rotating electric dipoles, eye blinks which generate electric charge variations (see
Figure 5, for examples of such artefacts). Such biosignals can of course provide
useful information (indicating discomfort, or attention lapses). One should be
aware of these artefacts and be careful not to process them indistinctly from
brain signals, for BCI that aims to process information coming strictly from the
brain [30].
11
Once an artefact has been detected, the portion of data corrupted with this
artefact should be rejected (i.e. disregarded) in two cases: 1) if the effect of the
artefact on the rest of the signal is too strong to be removed; 2) if the artefact
indicates that correct conditions are not met for pursuing with the BCI at that
particular time. In other cases, one can subtract the effect of the artefact from
the EEG by using special techniques such as Independent Components Analysis
or specific filters [57].
EEG signals can be represented in three different domains: in time (as they
are acquired by the amplifier), in frequency (spectral power) and in space (topographical layout on the scalp according to the electrode positions, as in Figure 3).
Time-based representations are especially useful when dealing with eventrelated potentials (ERPs). ERPs come with time triggers (see 2.3), which allow to extract epochs, which are time-windowed portions of signals. The EEG
measures a mixture of activities coming from many regions of the brain, and
although ERPs are stereotyped deflections, they are difficult to distinguish in
a single epoch. Statistical procedures are called for to extract the significant
deflections from several epochs (or trials). The simplest statistical procedure
is cross-trial averaging: this decreases the amplitude of the background oscillations, whose phase is not consistent across the trials. For Brain Computer
Interfaces where the analysis must be done in real time, there are specific techniques for single-trial detection of ERPs (see 3.2.2).
Frequency-based representations reveal the spectrum of the EEG, and
can be applied on sliding windows: such time-frequency representations are
necessary for BCI real-time processing. Since relevant brain activity is often
frequency-dependent (e.g. ERD/ERS reviewed in part 2.2), it is useful to bandpass filter the signals to the frequency-bands of interest. This also reduces the
influence of noise coming from other frequencies, e.g. power-line noise.
Spatial representations reveal spatial patterns, also called topographies,
related to the brain areas where the activity is produced. In order to enhance
the signal to noise ratio, some spatial processing can be applied to the EEG. Indeed, electric fields on the scalp result from electrical conduction within the head
volume. Because the skull is a poor electrical conductor, scalp topographies are
blurred compared to the distribution of cortical sources. Source reconstruction,
through an inverse problem, allows to pinpoint the activity of interest at the
cortical level [58]. BCI features can then be defined in source space [59, 60],
allowing to make use of a priori information. For instance, detection of an errorrelated potential is facilitated by only considering the source activity originating
from the anterior cingulate cortex. However, rather than source space analysis, simpler spatial filters, known as Laplacian filters, are more commonly used.
The Laplacian enhances the SNR of the data, by bringing out the differences
between neighboring electrodes. In fact, the Laplacian filter crudely approximates source reconstruction: it performs a deconvolution of the blurring due to
volume conduction [61, 62].
12
3.2
Feature extraction
As mentioned before, feature extraction aims at representing raw or preprocessed EEG signals by an ideally small number of relevant values, which describe
the task-relevant information contained in the signals. These features should
be selected to minimise the intra-class feature variances while maximising interclass variances. In other words, their values should be as different as possible
between different classes. BCI based on oscillatory activity (e.g., BCI based on
ERD/ERS or SSVEP) mostly use spectral and spatial information as features
whereas BCI based on ERP mostly use the temporal and spatial information.
The next sections detail the corresponding feature extraction tools for these two
categories of BCI.
3.2.1
13
Figure 6: Signal processing steps to extract band power features from raw EEG
signals. The EEG signal displayed here was recorded during right hand motor
imagery (the instruction to perform the imagination was provided at t = 0 s
on the plots), on electrode C3 (located over the left sensorimotor cortex), for 1
second. The contralateral ERD during imagination is here clearly visible in the
last half of the signal. Indeed, the signal power in channel C3 (left motor cortex)
in 8-12 Hz clearly decreases while the subject imagines a right hand movement.
x
(t) =
wi xi (t)
(1)
with x
(t) the spatially filtered signal, xi (t) the EEG signal from channel i
at time t and wi the weight given to that channel in the spatial filter. Spatial
filtering is not only useful because it reduces the dimensionality of the problem, but also because it can reduce spatial blur due to volume conduction as
mentioned in Section 3.1.
For BCI based on oscillatory EEG activity, the gold-standard in spatial filtering is a supervised algorithm (i.e. an algorithm that uses training data labelled
with their classes) known as the Common Spatial Patterns (CSP) algorithm
[68, 69]. Informally, the CSP algorithm finds spatial filters w that maximise
the variance of the filtered signal for one class and minimise it for the other
class. Since the variance of a band-pass signal is actually the band-power of this
signal, this means that CSP finds spatial filters that lead to maximally discriminant band-power features since their values are maximally different between
classes. In [70], it was shown that for two-class paradigms CSP maximizes an
approximation of mutual information of extracted EEG components and class
labels. As such, CSP is particularly useful for BCI based on oscillatory activity,
in which the relevant features are band-power features. As an example, for BCI
based on motor imagery, EEG signals are typically filtered in the 830 Hz band
before being spatially filtered with CSP [69]. Indeed this band contains both
14
the and rhythms. Formally, CSP uses the spatial filters w which extremise
(i.e. maximise or minimise) the following function:
J(w) =
wX1 X1T wT
wX2 X2T wT
(2)
wC1 wT
wC2 wT
(3)
(4)
i.e., the features used are simply the band power of the spatially filtered signals.
The use of CSP is illustrated in Figure 7. In this figure, the signals spatially
filtered with CSP clearly show differences in variance (i.e., in band power) between the two classes, hence ensuring high classification performances. It is
worth mentioning that the spatial filters w are only filters, and not the scalp topographies of the corresponding brain sources. As such the spatial filter weights
should not be interpreted from a neurophysiological point of view [71]. Indeed,
spatial filters can be seen as an unmixing or inverse model, whereas the forward
or mixing model is the one that should be neurophysiologically interpreted. To
go from the spatial filters (backward model) to the spatial patterns (forward
model), i.e., to the scalp topography of the corresponding brain sources, one
should work with the spatial patterns matrix A = (W 1 )T with W the matrix
with the w as columns [68, 71]. The spatial patterns, i.e., the columns of A, are
the scalp topographies of the brain sources targeted by the spatial filters W .
As mentioned above, CSP is a gold standard algorithm for BCI based on
oscillatory activity. Nonetheless, CSP has been shown not to be very robust
to noise, non-stationarities and artefacts and to be prone to overfitting (i.e.,
15
Figure 7: EEG signals spatially filtered using the CSP (Common Spatial Patterns) algorithm. The first spatial filter (top filter) maximises the variance of
signals from class Left Hand Motor Imagery while minimising that of class
Right Hand Motor Imagery. It corresponds to the largest eigenvalue of the
GEVD. The last filter (bottom filter) is the opposite: it maximises the variance
of class Right Hand Motor Imagery while minimising that of class Left Hand
Motor Imagery (it corresponds to the lowest eigenvalue of the GEVD). This
can be clearly seen during the periods of right or left hand motor imagery, in
green and red respectively.
it may not generalise well to new data) when little training data is available
[70, 72, 73]. As such, many CSP variants have been designed to make it more
robust to noise and non-stationarities [74, 75], or to enable it to optimise both
spatial and spectral filters at the same time [76, 77].
3.2.2
Like BCI based on oscillatory activity, BCI based on ERP can exploit the spatial
information in EEG signals. This is typically done by focusing mostly on some
specific electrodes (i.e., by extracting features only for these electrodes), e.g.
for the P300, by using only electrodes located over the parietal lobe, where the
P300 is known to originate. As an example, Krusienski et al. recommend to
use a set of 8 channels, in positions Fz, Cz, P3, Pz, P4, PO7, Oz, PO8 [78].
Once the relevant spatial information has been identified, using, for instance,
only the electrodes mentioned above, features can be extracted for the signal of
each of them. For ERP in general, including the P300, the features generally
exploit the temporal information of the signals, i.e., how the amplitude of the
EEG signal varies with time. This is typically achieved by using the values of
preprocessed EEG time points as features. More precisely, features for ERP
are generally extracted by 1) low-pass or band-pass filtering the signals (e.g., in
1-12 Hz for the P300), 2) downsampling the filtered signals, in order to reduce
the number of EEG time points and thus the dimensionality of the problem and
3) gathering the values of the resulting EEG time points from all considered
channels into a feature vector that will be used as input to a classifier. This
process is illustrated in Figure 8, for feature extraction from channel Pz in a
P300-based BCI experiment.
With ERPs the number of features is usually quite large, with many features
16
Figure 8: Typical process to extract features from a channel of EEG data for
a P300-based BCI design. On this picture we can see the P300 becoming more
visible with the different processing steps.
per channel and many channels used. As such, tools such as feature selection
or spatial filtering can also benefit ERP-based BCI design. Regarding spatial
filtering, some groups have proposed spatial filters dedicated to the classification
of ERPs, leading to maximally discriminant ERP features [79, 80]. One of them
is the xDAWN spatial filter [80]. This spatial filter aims at maximising the
signal-to-signal-plus-noise-ratio. Informally, this means that xDAWN aims at
enhancing the ERP response, at making the ERP more visible in the middle of
the noise. Formally, xDAWN finds spatial filters that maximise the following
objective function:
wADDT AT wT
(5)
wXX T wT
where A is the time course of the ERP response to detect for each channel
(estimated from data, usually using a Least Square estimate) and D is a matrix containing the time triggers of the target stimuli evoking the ERP. In this
equation, the numerator represents the signal, i.e., the relevant information to
enhance. Indeed, wADDT AT wT is the power of the time course of the ERP responses after spatial filtering. On the contrary, in the denominator, wXX T wT
is the variance of all EEG signals after spatial filtering. Thus, it contains both
the signal (the ERP) plus the noise. Therefore, maximising JxDAWN actually
maximises the signal, i.e., it enhances the ERP response, and simultaneously
minimises the signal plus the noise, i.e., it makes the noise as small as possible
[80]. This has indeed been shown to lead to much better ERP classification
performance, especially when little training data was available.
JxDAWN =
17
3.3
Pattern recognition
The pattern recognition stage aims at predicting a value from which infering the
users current brain activity pattern. It uses as input some features or patterns
extracted from brain signals such as the power in specific frequency bands or a
smoothed signal segment (see 3.2). Predicting a discrete variable corresponds
to a classification task while predicting a continuous variable corresponds to a
regression task.
Classification consists first in learning an efficient function f from a training
set of N labelled examples {(d1 , c1 ), (d2 , c2 ), ..., (dN , cN )} where dn is a feature
vector and cn is its associated class, to allow to assign a new feature vector d to
a unique class c = f (d) among C. When C only has 2 elements, the classification is binary. Numerous machine learning techniques, such as support vector
machines, neural networks or logistic function, can perform this task. If C
contains more than 2 classes, there are two possible approaches to obtain a classification function [81, 82]. The first approach consists in directly estimating the
class using multiclass techniques such as decision trees, multilayer perceptrons,
naive Bayes classifiers or k-nearest neighbours. The second approach consists
in decomposing the problem into several binary classification problems. This
decomposition can be done in different ways using i) one-against-one pairwise
classifiers [83, 84], ii) one-against-the rest (or one-against-all) classifiers [84], iii)
hierarchical classifiers similar to a binary decision tree and iv) multilabel classifiers [85, 84]. In the later case, a distinct subset of L labels (or properties)
is associated to each class. The predicted class is identified according to the
closest distance between the predicted labels and each subset of labels defining
a class.
The performance of the classification function f can be evaluated using different statistical measures. The most popular measure is the accuracy which
measures the percentage of correct recognition. The accuracy is sensitive to the
number of classes. The Cohens Kappa coefficient is a performance measure
which takes the number of classes into consideration. However, these measures
do not distinguish types of error. Indeed, when the classifier assigns a feature
vector (or pattern) to class A whereas it actually belongs to class B, this error
has the same cost as the opposite error. In many applications, these two types
of error (namely type I and II in statistical hypothesis testing) do not have the
same impact. As an example, if the detection of an ERP triggers the writing
of a letter, it can be more acceptable for the system not to write a letter but
instead to wait for more data in order not to write an erroneous letter.
To get a detailed measure of types of error, it is possible to build a confusion
matrix. This 2D-table reports how many patterns from each class (true label)
have been assigned to which class (estimated label). The occurrences in the diagonal correspond to the correct assignments. Occurrences outside the diagonal
correspond to different types of errors. If it is more important to take care of the
performance of a specific class (named P as positive), it is possible to compute
the recall (or sensitivity) and the precision of the classification function for this
18
class. Correctly classified patterns of this class are called True Positives (TP)
because the specific class to look at is identified as the positive one, and the
other classes are identified as the negative ones. The recall is computed as the
ratio of TP to the total number of patterns of class P. This measure indicates
how good the function is at identifying patterns of class P. Nevertheless, a good
value can be due to a biased classifier which excessively assigns any pattern to
class P. So, it is also useful to measure the precision which corresponds to the
ratio of TP to the number of patterns assigned to class P. Patterns incorrectly
assigned to class P are called False Positives (FP). Patterns of class P assigned
to another class are called False Negative (FN) and patterns of other classes
than P not assigned to class P are called True Negatives (TN). In theory, both
the recall and the precision rates are useful. The F-score combines these two
measures in one.
Here is a summary of the statistical measures presented above:
accuracy = (T P + T N )/(T P + F P + F N + T N );
= (accuracy - probability of random agreement)/(1 - probability of random agreement);
recall (or sensitivitxy) = T P/(T P + F N ) (or true positive rate);
specificity = T N/(F P + T N ) (or true negative rate);
precision = T P/(T P + F P );
F-score =
2recallprecision
recall+precision .
In practice, it is unfortunately frequent that the recall and the precision cannot
both be close to one. Thus, according to the application, it is sometimes better
to favour a high recall (e.g. detection of ErrP, see section 2.3) or a high precision (e.g. detection of ERP). The Receiver Operating Characteristic (ROC)
curve is used to observe the effect of the decision threshold on the sensitivity
and specificity. The surface of this curve is also used as an indicator to evaluate
classifier performance and to compare classifiers. Indeed, many different machine learning techniques exist for classification. The most used in BCI are the
Linear Discriminant Analysis (LDA) and the Support Vector Machines (SVM),
for their capabilities to obtain a robust classifier i.e. efficient with new patterns. LDAs are simple linear models which do not require any hyperparameter
[20]. SVM are linear or non-linear models, depending on the selected kernel,
that maximise the distance between the decision boundary and the patterns of
each class [20]. Robustness, or generalisation capability, is an important issue.
The classification function learned during the calibration phase must be efficient with new patterns. Thus, the classification function is evaluated using the
statistical measures presented above on a subset of labelled patterns acquired
for calibration. If the dataset is large, the evaluation can be done using only
one subset. When it is small, a cross-validation procedure is used for which the
labelled dataset is split into N folds (e.g. N = 10) [86]. Then the parameters
19
3.4
Feedback
Designing a reliable BCI requires to consider its two main components, the user
and the machine. BCI should be seen as a co-adaptive system [6], with the
user learning to produce EEG patterns that can be recognised by the machine,
and the machine learning to recognise such patterns. Indeed, BCI control is
known to be a skill that needs to be learned and mastered by the user [6, 22].
20
This is notably essential for spontaneous BCI, that is, BCI based on spontaneously generated brain activity patterns (e.g., motor imagery-based BCI).
Indeed, ERP-based BCI require very little user training [92]. The fact that
spontaneous BCI control is a skill means that 1) the BCI performances of a
user become better with practice and thus that 2) the user needs to learn how
to produce stable, clear and distinct brain activity patterns to successfully control a BCI. A key and necessary element for successful learning of spontaneous
BCI control skills is the feedback provided by the machine to the user about
his/her brain activity, i.e., about the mental command recognised by the BCI
[22, 93, 94]. Typically, this feedback takes the form of a visual cursor or a bar
moving on screen according to the recognised brain activity pattern (see Figure
9).
research efforts to design an ideal BCI feedback and training environment [33].
Applications
Once the BCI is able to recognise the users brain activity patterns, it can be
used to send commands to an application. Below we present some of the main
applications in which BCI can be used.
4.1
Spellers
Several BCI systems, called spellers have been proposed for typing sentences
without muscular control, thus eliciting communication for severely disabled
patients. The most successful of these spellers is called the P300 speller [10].
The P300 speller has the advantage not to require any training on the part of
the user, whose task is to concentrate his/her attention on the symbol he/she
wants to type. The P300 speller can predict the symbol on which the users
attention is focussing, because flashing this symbol elicits an ERP in the EEG,
the so-called P300 wave. Figure 10 illustrates the P300 speller principle.
Figure 10: This user wants to type the letter O: he counts the number of
times it flashes on the screen, as in the lower picture. Each flash of the target
letter creates an Evoked Response Potential, the P300 wave, time-locked to the
flash display.
The BCI system has to be calibrated to detect the P300 wave from the
background EEG (see section 2.3). To accelerate the throughput, keyboard
symbols are organized in groups that flash simultaneously. The original P300
had a row-column flashing strategy [10], but groups can be arranged in other
patterns, e.g. to minimize the probability of a given symbol flashing twice in
22
a row [95]. Indeed, double flashes are both uncomfortable for the user, and
produce a reduced P300 evoked response for the second flash.
The way a symbol is flashed generally consists of increasing its visual contrast
with the background, and also increasing its size. Recent findings by Jin et al.
show that the P300 signal-to-noise can be increased by replacing the symbols
with pictures such as famous faces during the flashing interval [96].
The signal-to-noise ratio of a P300 wave is low, so sequences of flashing symbols must be repeated to make its detection possible. Each group of symbols
is generally flashed several times. Besides, users are generally asked to count
the number of flashes of their desired symbol in order to engage their attention. Modern P300 systems now use early stopping of the flashes as soon as a
reliable decision has been made [97]. Early stopping increases the throughput,
while improving users motivation, whose focussed attention is rewarded by a
faster decision [98, 99]. In spite of the progress made in this application, it can
take several seconds for a user to type a symbol with the P300 speller. Word
prediction engines can be combined with the speller, to make the typing more
efficient [100].
Of course, the P300 spellers can be adapted to any task relying on a choice
among several elements. BCI games have been developed using this principle
(see section 4.5).
A few other BCI spellers are based on other neurophysiological features than
the P300. The commercial Intendix system http://www.intendix.com/ can be
operated either with a P300 or an SSVEP. The Hex-O-Spell [101] uses self-paced
MI features to control a rotating arrow, allowing to point to the desired letter
(in two steps, each with only 6 choices).
Flashing, or even displaying symbols on a screen may not be appropriate
for all users, say in the case of visual deficiencies [102]. Auditory P300 spellers
have been developed,whose throughput is generally lower than for visual P300
spellers, because less symbols can be presented simultaneously to the user [103].
When hardware becomes easier to handle and to set up, BCI spellers could
become commonly used by individuals with severe neuromuscular disabilities,
such as amyotrophic lateral sclerosis (ALS). Several clinical studies have indeed demonstrated the usability of the spellers [104, 99], for patients who are
able to maintain a sustained attention [105], and regular home use has been
reported [106].
4.2
Motor substitution
the nerve fibers. The activity of the motor cortex is then analysed to decode
the intention of movement. Using multielectrode arrays placed invasively in
the primary motor cortex, it is possible to decode the direction of movement
from the activity of a neural population [108] and send commands to a robotic
limb or an exoskeleton to reach and grasp a target [109]. Electrocorticography
electrodes placed on the cortex are also used to obtain a good enough quality of
signals, less invasively than microelectrode arrays which can cause brain tissue
damage. Non-invasive signals such as EEG are also used in neuroprosthetics
[12] although the spatial resolution is low and the signals are noisy. Thus,
P300 wave and SSVEP can be used for a control task [50]. But ERD is most
commonly used to detect a desynchronisation in various parts of the cortex
meaning that a specific mental task is performed. As an example, each mental
task can be translated into a command for a wheelchair in a shared-control
with an obstacle avoidance system [110, 111]. Motor imagery tasks are often
used to control neural prosthesis or to trigger a functional electrical stimulation
system [112, 113]. It should be noted, however, that designing neuroprostheses
for the lower limb is still a challenge [114], notably due to the lack of tactile
or proprioceptive feedback to the user. Therefore, current efforts are focussed
on adding stimulations in the somatosensory cortex or on the skin of user to
provide him/her with tactile feedback and therefore to obtain a complete actionperception loop.
4.3
Stroke rehabilitation
4.4
4.5
As mentioned earlier, potential BCI applications are far from only targeting
severely motor impaired users, and can also benefit healthy users [127]. BCI can
notably be seen as an exciting alternative or complementary way to control video
games [128, 14, 129] and/or other virtual reality (VR) applications [13, 130].
For instance, BCI can enable a gamer to select objects in the game simply by
paying attention to them [131, 132, 133]. In this case, a specific visual stimulus is
displayed in overlay of each object, in order to evoke P300 or SSVEP responses.
For instance, a BCI version of the popular connect-4 game was proposed
based on the P300 response [134]. In this game, the players could select the
column in which to insert their token thanks to P300-based selection of the
flashing columns. BCI can also be used to perform navigation tasks in game
virtual environments VE [135, 136, 137, 138, 139]. This can be achieved for
instance by performing left hand, right hand and foot MI to turn left, right
and go forward respectively [135, 139, 136] or by paying attention to flickering
control buttons overlayed to the game, each button evoking a different SSVEP
response and being associated to a different displacement direction in the game
[138, 140]. In all these BCI-based video games prototypes, the BCI was the sole
input device. However, current BCI typically provide a rather limited number of
commands (typically between 1 and 3 - with substantially lower accuracy with
higher number of commands), and thus provide a limited interaction with the
game. An interesting approach is therefore to use the BCI as a complementary
25
input device: most of the game is controlled using a traditional input device such
as a game pad, and only some actions, e.g., special powers, are controlled using
the BCI. This was explored in the alphaWoW (alpha World of Warcraft) game,
in which the players avatar in the WoW game was controlled using a game
pad, and this avatar can turn from a brutal bear into a fragile elf depending on
the players relaxation level, as measured in his/her EEG signals in the alpha
band [141].
4.6
Most, if not all, of the BCI-based systems and applications described so far
were systems for communication and control, in which the user was voluntarily
sending mental commands to the application. These types of BCI are known
either as active BCI, when the user performs mental tasks (e.g., motor imagery),
or as reactive BCI, when the users have to attend to stimuli (e.g., P300 or
SSVEP-BCI) [15]. There is yet another category of BCI: passive BCI, for which
the mental state of the user is passively estimated, without any voluntary mental
command from the user, in order to adapt the application in real-time to this
mental state [15]. For instance, some passive BCI can be used to monitor
vigilance/alertness levels of operators performing repetitive tasks, such as x-ray
luggage controllers in airports [142]. Detecting lapses of attention reduces the
risk of missing threats, and can be used to propose a break to tired operators.
Another example would be a passive BCI used in a car, to detect emergency
braking intentions. Emergency braking intentions can indeed be detected in
EEG signals before the actual physical braking, hence making it possible to
stop the car earlier, potentially avoiding an accident [143]. Note that in this
specific case, the used BCI design and signal processing approach are similar
to that of active BCI based on movement onset detections. However, in the
case of emergency braking intentions, the user does not have to do anything
specific to control the BCI: he/she is driving his/her car normally, while his/her
brain activity is passively monitored to detect emergency braking intention.
There are actually increasingly more passive BCI application prototypes being
proposed, and which were reviewed in [144, 15]. Indeed, a number of interesting
mental states can be estimated, more-or-less reliably, in EEG signals, such as
mental workload (i.e., pressure on the user cognitive resources, e.g., working
memory) [145, 146], attention [142, 147] and, though less accurately, emotions
[148], among others [149]. Being able to monitor continuously such mental
states from EEG signals makes it possible to design applications that adapt to
such mental states, to provide an improved user experience [150] or simply to
assess the ergonomic qualities of the system [151, 152], among many potential
applications.
26
27
Overall, EEG-based BCI is currently a very active and dynamic research topic,
with hundreds of labs involved, and with many promising potential outcomes
in the future. Nonetheless, EEG-based BCI are still far from being free of limitations. In particular, they still suffer for a relatively poor reliability [28, 6], as
they often erroneously interpret the users mental commands, and still do not
work for a substantial portion of users. This may prevent them from competing with alternative input devices, e.g., eye-trackers for motor impaired users
who retain some eye movements. The usability of BCI must still be improved,
with more comfortable and practical EEG caps and electrodes, shorter calibration, and increased support for operating and exploring these systems which are
generally not plug-and-play. The reliability issue, long calibration times and
lack of comfort are issues that currently prevent BCI from actually reaching
the mainstream video game market, as the gamer population is very has high
expectations regarding speed, accuracy and ease-of-use of their gaming devices
[163].
Many research efforts are still needed to bring EEG-based BCI outside laboratories, to users homes, to the consumer market and to hospitals, for daily
use. To improve the reliability of BCI, some interesting research directions are
a) to improve machine learning algorithms by collecting very large data sets
of EEG signals from multiple contexts and users, to improve BCI calibrations;
b) to explore new developments in machine learning and signal processing in
order to design invariant features and classifiers that could potentially greatly
increase the robustness and reliability of the system [164]; and c) to improve
BCI user training protocols to make them comply with human learning theories
and guidelines, hence hopefully dramatically increasing users BCI control skills
[33]. Improvement in neuroscientific knowledge could also lead to the discovery
of new features or new brain activity patterns than can be exploited reliably for
BCI. Conversely, BCI opens the way to new neuroscientific experiments with
the user in the loop [165].
With BCI becoming more mature, and increasingly more brain activity patterns being detectable in EEG signals, many new BCI-based applications could
and should be explored. This is notably the case of passive BCI, which is
still rather recent and thus still underexplored but very promising, for healthy
users as well as for patients (e.g., for patients mental states monitoring and
automatic treatment). New applications of BCI could further boost R&D in
the area. According to a recent questionnaire sent to the BCI community, the
coordination group of the BNCI Horizon 2020 project identified four visions
[166]: i) Patients will use invasive BCIs to compensate for movement disorders; ii) Passive BCIs will enrich human-computer interaction; iii) BCIs will be
commonly used as tools for basic research and iv) the use of BCI technology
for rehabilitation will increase. For six application scenarios (replace, restore,
enhance, supplement, improve, research tool [6]), a future use case has been reported (see http://bnci-horizon-2020.eu/roadmap) respectively: unlocking
the completely locked-in, neuroprosthesis, user experience in computer games,
28
BCI-controlled glass, upper limb rehabilitation after stroke, tools for cognitive
neurosciences. The associated roadmap specify for the six application scenarios
the available knowledge in 2015, the research in progress up to 2025 and the
possible state-of-the-art in 2035 (resp. speech decoder, decoding of complex
movements, wearable devices & applications, BCIs for rehabilitation at home,
plug and play tools for research).
When developing BCI, it is also essential to consider the underlying ethical
issues, regarding the brain activity patterns to measure and to use (e.g., should
we access and reveal them? and what for?) or about how, when and whether
to use them with motor impaired or healthy users (e.g., should BCI be used
to enhance an individuals abilities?), among others [167]. Overall, EEG-based
BCI research has very exciting research and development perspectives in the
years to come.
References
[1] H. Berger. Ueber das elektroenkephalogramm des menschen. Archiv fur
Psychiatrie und Nervenkrankheiten, 87:527570, 1929.
[2] Alvin Lucier. Music for solo performer. Lovely Music, Ltd., 1982. for
enormously amplified brain waves and percussion - originally composed in
1965.
[3] Pierre Henry. Mise en musique du corticalart de Roger Lafosse. Philips
Prospective 21e sicle, 1971. Vinyl LP.
[4] J. J. Vidal. Toward direct brain-computer communication. Annual Review
of Biophysics and Bioengineering, 2:157180, 1973.
[5] J.R. Wolpaw, N. Birbaumer, D.J. McFarland, G. Pfurtscheller, and T.M.
Vaughan. Brain-computer interfaces for communication and control. Clinical Neurophysiology, 113(6):767791, 2002.
[6] J. R. Wolpaw and E. W. Wolpaw. BCI: Principles and Practice. Oxford
University Press, 2012.
[7] J. R. Wolpaw, D. J. McFarland, G. W. Neat, and C. A. Forneris. An EEGbased brain-computer interface for cursor control. Electroencephalography
and clinical neurophysiology, 78:252259, 1991.
[8] G. Pfurtscheller, D. Flotzinger, and J. Kalcher. Brain-computer interfacea new communication device for handicapped persons. journal of microcomputer application, 16:293299, 1993.
[9] N. Birbaumer, N. Ghanayim, T. Hinterberger, I. Iversen, B. Kotchoubey,
A. K
ubler, J. Perelmouter, E. Taub, and H. Flor. A spelling device for
the paralysed. Nature, 398:297298, 1999.
29
[10] L.A. Farwell and E. Donchin. Talking off the top of your head: toward a mental prosthesis utilizing event-related brainpotentials. Electroencephalography and Clinical Neurophysiology, 70(6):510 523, 1988.
[11] Bernhard Graimann, Brendan Allison, and Gert Pfurtscheller. Braincomputer interfaces:
Revolutionizing human-computer interaction.
Springer, 2010.
[12] J.R. Mill
an, R. Rupp, G. M
uller-Putz, R. Murray-Smith, C. Giugliemma,
M. Tangermann, C. Vidaurre, F. Cincotti, A. K
ubler, R. Leeb, C. Neuper, K.-R. M
uller, and D. Mattia. Combining brain-computer interfaces
and assistive technologies: State-of-the-art and challenges. Frontiers in
Neuroprosthetics, 2010.
[13] A. Lecuyer, F. Lotte, R.B. Reilly, R. Leeb, M. Hirose, and M. Slater.
Brain-computer interfaces, virtual reality and videogames. IEEE Computer, 41(10):6672, 2008.
[14] D Coyle, J Principe, F Lotte, and A Nijholt.
Guest editorial:
Brain/neuronal computer games interfaces and interaction. IEEE Transactions on Computational Intelligence and AI in Games, 5(2):7781, 2013.
[15] T.O. Zander and C. Kothe. Towards passive brain-computer interfaces:
applying brain-computer interface technology to human-machine systems
in general. Journal of Neural Engineering, 8, 2011.
[16] J.J. Daly and J.R. Wolpaw. Brain-computer interfaces in neurological
rehabilitation. The Lancet Neurology, 7(11):10321043, 2008.
[17] Kai Keng Ang and Cuntai Guan. Brain-computer interface in stroke rehabilitation. Journal of Computing Science and Engineering, 7(2):139146,
2013.
[18] J.R. Wolpaw, G.E. Loeb, B.Z. Allison, E. Donchin, O.F. do Nascimento,
W.J. Heetderks, F. Nijboer, W.G. Shain, and J. N. Turner. BCI meeting
2005workshop on signals and recording methods. IEEE Transaction on
Neural Systems and Rehabilitation Engineering., 14(2):138141, 2006.
[19] A. Bashashati, M. Fatourechi, R. K. Ward, and G. E. Birch. A survey of
signal processing algorithms in brain-computer interfaces based on electrical brain signals. Journal of Neural engineering, 4(2):R3557, 2007.
[20] F. Lotte, M. Congedo, A. Lecuyer, F. Lamarche, and B. Arnaldi. A review of classification algorithms for EEG-based brain-computer interfaces.
Journal of Neural Engineering, 4:R1R13, 2007.
[21] A. K
ubler, V.K. Mushahwar, L.R. Hochberg, and J.P. Donoghue. BCI
meeting 2005-workshop on clinical issues and applications. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 14(2):131134,
2006.
30
[22] C. Neuper and G. Pfurtscheller. Brain-Computer Interfaces, chapter Neurofeedback Training for BCI Control, pages 6578. The Frontiers Collection, 2010.
[23] Pieter-Jan Kindermans, Michael Tangermann, Klaus-Robert M
uller, and
Benjamin Schrauwen. Integrating dynamic stopping, transfer learning and
language models in an adaptivezero-training erp speller. Journal of Neural
Engineering, 11(3):035005, 2014.
[24] F. Lotte. Signal processing approaches to minimize or suppress calibration
time in oscillatory activity-based brain-computer interfaces. Proceedings
of the IEEE, 2015.
[25] C. Guger, G. Edlinger, W. Harkam, I. Niedermayer, and G. Pfurtscheller.
How many people are able to operate an EEG-based brain-computer interface (BCI)? IEEE Transactions on Neural Systems and Rehabilitation
Engineering, 11(2):145147, 2003.
[26] C. Guger, S. Daban, E. Sellers, C. Holzner, G. Krausz, R. Carabalona,
F. Gramatica, and G. Edlinger. How many people are able to control a P300-based brain-computer interface (BCI)? Neuroscience Letters,
462(1):9498, 2009.
[27] Christoph Guger, Brendan Z Allison, Bernhard Growindhager, Robert
Pr
uckl, Christoph Hinterm
uller, Christoph Kapeller, Markus Bruckner,
Gunther Krausz, and G
unter Edlinger. How many people could use an
ssvep bci? Frontiers in neuroscience, 6, 2012.
[28] B.Z. Allison and C. Neuper. Could anyone use a BCI? In Desney S.
Tan and Anton Nijholt, editors, Brain-Computer Interfaces, volume 0 of
Human-Computer Interaction Series, pages 3554. Springer London, 2010.
[29] S. Brandl, J. H
ohne, K.-R. M
uller, and W Samek. Bringing bci into everyday life: Motor imagery in a pseudo realistic environment. In roceedings
of the International IEEE/EMBS Neural Engineering Conference (NER),
2015.
[30] M. Fatourechi, A. Bashashati, R. Ward, and G. Birch. EMG and EOG
artifacts in brain computer interface systems: A survey. Clinical Neurophysiology, 118(3):480494, 2007.
[31] F. Lotte, J. Fujisawa, H. Touyama, R. Ito, M. Hirose, and A. Lecuyer.
Towards ambulatory brain-computer interfaces: A pilot study with P300
signals. In 5th Advances in Computer Entertainment Technology Conference (ACE), pages 336339, 2009.
[32] N. Neumann and A. Kubler. Training locked-in patients: a challenge
for the use of brain-computer interfaces. IEEE Transactions on Neural
Systems and Rehabilitation Engineering, 11(2):169172, 2003.
31
[33] Fabien Lotte and Camille Jeunet. Towards improved bci based on human learning principles. In 3rd International Brain-Computer Interfaces
Winter Conference, 2015.
[34] E. Niedermeyer and F. Lopes da Silva. Electroencephalography: Basic
Principles, Clinical Applications, and Related Fields. Philadelphia , Lippincott Williams & Wilkins, 2005.
[35] H. H. Jasper. The ten-twenty electrode system of the international federation. Electroencephalography and Clinical Neurophysiology, 10(2):371375,
1958.
[36] G. Pfurtscheller and F. H. Lopes da Silva. Event-related EEG/MEG
synchronization and desynchronization: basic principles. Clinical Neurophysiology, 110(11):18421857, 1999.
[37] G. Pfurtscheller and C. Neuper. Motor imagery and direct brain-computer
communication. proceedings of the IEEE, 89(7):11231134, 2001.
[38] E.V.C. Friedrich, R. Scherer, and C. Neuper. Long-term evaluation of a
4-class imagery-based braincomputer interface. Clinical Neurophysiology,
124(5):916 927, 2013.
[39] J.R. Mill
an, J. Mouri no, M. Franze, F. Cincotti, M. Varsta, J. Heikkonen,
and F. Babiloni. A local neural classifier for the recognition of eeg patterns associated to mental tasks. IEEE transactions on neural networks,
13(3):678686, 2002.
[40] J. Kronegg, G. Chanel, S. Voloshynovskiy, and T. Pun. EEG-based synchronized brain-computer interfaces: A model for optimizing the number
of mental tasks. IEEE Transactions on Neural Systems and Rehabilitation
Engineering, 15(1):5058, 2007.
[41] Connie C. Duncan, Robert J. Barry, John F. Connolly, Catherine Fischer, Patricia T.Michie, Risto Naatanen, John Polich, Ivar Reinvang, and
Cyma Van Petten. Event-related potentials in clinical research: Guidelines for eliciting, recording, andquantifying mismatch negativity, P300,
and N400. Clinical Neurophysiology, 120(11):1883 1908, 2009.
[42] R.M. Chapman and H.R. Bragdon. Evoked responses to numerical and
non-numerical visual stimuli while problem solving. Nature, 203:115557,
1964.
[43] John Polich. Updating p300: An integrative theory of p3a and p3b. Clinical Neurophysiology, 118(10):2128 2148, 2007.
[44] Ricardo Chavarriaga, Aleksander Sobolewski, and Jose del R Millan. Errare machinale est: The use of error-related potentials in brain-machine
interfaces. Frontiers in Neuroscience, 8(208), 2014.
32
33
35
38
L
aer, Ozge
Yilmaz, Fabricio L Brasil, Giulia Liberati, Marco R Curado,
Eliana Garcia-Cossio, Alexandros Vyziotis, et al. Brainmachine interface
in chronic stroke rehabilitation: A controlled study. Annals of neurology,
74(1):100108, 2013.
[120] Donatella Mattia, Floriana Pichiorri, Marco Molinari, and R
udiger Rupp.
Brain computer interface for hand motor function restoration and rehabilitation. In Towards Practical Brain-Computer Interfaces, pages 131153.
Springer, 2013.
[121] Kai Keng Ang, Cuntai Guan, Kok Soon Phua, Chuanchu Wang,
Longjiang Zhou, Ka Yin Tang, Gopal Joseph Ephraim Joseph, Christopher Wee Keong Kuah, and Karen Sui Geok Chua. Brain-computer
interface-based robotic end effector system for wrist and hand rehabilitation: results of a three-armed randomized controlled trial for chronic
stroke. Frontiers in Neuroengineering, 7(30), 2014.
39
[122] Damien Coyle, Jacqueline Stow, Karl McCreadie, Jacinta McElligott, and
Aine
Carroll. Sensorimotor modulation assessment and brain-computer interface training in disorders of consciousness. Archives of physical medicine
and rehabilitation, 96(3):S62S70, 2015.
[123] Dorothee Lule, Quentin Noirhomme, Sonja C Kleih, Camille Chatelle, Sebastian Halder, Athena Demertzi, Marie-Aurelie Bruno, Olivia Gosseries,
Audrey Vanhaudenhuyse, Caroline Schnakers, et al. Probing command
following in patients with disorders of consciousness using a brain
computer interface. Clinical Neurophysiology, 124(1):101106, 2013.
[124] Gernot R M
uller-Putz, Christoph Pokorny, Daniela S Klobassa, and Petar
Horki. A single-switch bci based on passive and imagined movements:
toward restoring communication in minimally conscious patients. International journal of neural systems, 23(02), 2013.
[125] Camille Chatelle, Srivas Chennu, Quentin Noirhomme, Damian Cruse,
Adrian M Owen, and Steven Laureys. Brain-computer interfacing in disorders of consciousness. Brain injury, 26(12):15101522, 2012.
[126] Jacques Luaute, Dominique Morlet, and Jeremie Mattout. Bci in patients
with disorders of consciousness: Clinical perspectives. Annals of Physical
and Rehabilitation Medicine, 2015.
[127] J. van Erp, F. Lotte, and M. Tangermann. Brain-computer interfaces:
Beyond medical applications. IEEE Computer, 45(4):2634, 2012.
[128] A. Nijholt, D. Plass-Oude Bos, and B. Reuderink. Turning shortcomings into challenges: Brain-computer interfaces for games. Entertainment
Computing, 1(2):8594, 2009.
[129] David Marshall, Damien Coyle, Shane Wilson, and Michael Callaghan.
Games, gameplay, and bci: The state of the art. Computational Intelligence and AI in Games, IEEE Transactions on, 5(2):8299, 2013.
[130] Fabien Lotte, Josef Faller, Christoph Guger, Yann Renard, Gert
Pfurtscheller, Anatole Lecuyer, and Robert Leeb. Combining BCI with
Virtual Reality: Towards new applications and improved BCI. In Brendan Z. Allison, Stephen Dunne, Robert Leeb, Jose Del R. Millan, and
Anton Nijholt, editors, Towards Practical Brain-Computer Interfaces, Biological and Medical Physics, Biomedical Engineering, pages 197220.
Springer Berlin Heidelberg, 2013.
[131] J. D. Bayliss. The use of the P3 evoked potential component for control in a virtual apartment. IEEE Transactions on Neural Systems and
Rehabilitation Engineering, 11(2):113116, 2003.
[132] E. Lalor, S. P. Kelly, C. Finucane, R. Burke, R. Smith, R. Reilly, and
G. McDarby. Steady-state VEP-based brain-computer interface control
40
[137] F. Velasco-Alvarez
and R. Ron-Angevin. Free virtual navigation using motor imagery through an asynchronous brain-computer interface. Presence:
teleoperators and virtual environments, 19(1):7181, 2010.
[138] J. Faller, G. M
uller-Putz, D. Schmalstieg, and G. Pfurtscheller. An application framework for controlling an avatar in a desktop-based virtual
environment via a software SSVEP brain-computer interface. Presence:
teleoperators and virtual environments, 19(1):2534, 2010.
[139] F. Lotte, A. Van Langhenhove, F. Lamarche, T. Ernest, Y. Renard, B. Arnaldi, and A. Lecuyer. Exploring large virtual environments by thoughts
using a brain-computer interface based on motor imagery and high-level
commands. Presence: teleoperators and virtual environments, 19(1):54
70, 2010.
[140] J Legeny, R Viciana Abad, and A Lecuyer. Toward contextual ssvep-based
bci controller: Smart activation of stimuli and controls weighting. IEEE
Transaction on Computational Intelligence and AI in Games, 2013.
[141] Bram van de Laar, Hayrettin Gurkok, D Plass-Oude Bos, Mannes Poel,
and Anton Nijholt. Experiencing bci control in a popular computer game.
Computational Intelligence and AI in Games, IEEE Transactions on,
5(2):176184, 2013.
[142] B. Blankertz, M.Tangermann, C. Vidaurre, S. Fazli, C.Sannelli, S.Haufe,
C. Maeder, L. Ramsey, I. Sturm, G. Curio, and K.R. Mueller. The berlin
41
In
[164] DJ Krusienski, M Grosse-Wentrup, F Galan, D Coyle, KJ Miller, E Forney, and CW Anderson. Critical issues in state-of-the-art braincomputer
interface signal processing. Journal of Neural Engineering, 8(2):025002,
2011.
43
44